A number without its configuration is a marketing claim, not evidence. Every result we publish carries its model, hardware, context length, and the seed it can be replayed from, measured against the same workload on a standard cache. Anything that has not cleared that bar is not on this site.
A footprint or a recall delta means nothing on its own. Read on the wrong hardware, at the wrong context length, against a baseline you never saw, it can say almost anything. So we never report a value alone.
Each result ships as a card: the metric, the configuration that produced it, the seed it replays from, and the baseline it is measured against. Change one field and it is a different result, reported separately.
Illustrative schema. The fields are fixed, the values come from the run.
The total field budget under a stated configuration, shown against a standard cache on the same workload, so the ceiling you chose is the ceiling you get.
Attention to early-conversation content, measured across the run rather than asserted at the end of it. Reported against the same workload on a standard cache, at a stated context length, so the comparison is a measurement rather than a claim.
Read and write latency at each fidelity setting, on the hardware named in the result, so the trade is visible before you commit to it.
The conditions where SGF underperforms, published next to the wins, because a number you can trust is one that admits its own edges.
SCOPE: these are the axes. The published run above reports footprint and validity; recall and latency for your own workload come out of the pilot, each carrying the configuration that produced it.
The card on the right is a real result, not a format demonstration. One unbroken conversation of 1,900,000 tokens across 9,359 turns, on an unmodified off-the-shelf model, with the field footprint unchanged from 50,000 tokens onward. The run is measured live and retained in full; its seed fixes the input workload, so the same conversation can be measured against a standard cache. The scope note below states what the card does and does not claim.
Measured values. Total deployed memory across this run was approximately 45 MB against approximately 62 GB for a standard KV cache at the same depth. The 19.1 MB above is the cache replacement and is flat across the run; total deployed memory stays bounded in the low tens of MB at any depth. The ratio grows with length and is stated at its depth. This run demonstrates topical coherence and continued attention to early-conversation content. It is not a claim of exact fact recall. The field runs on a single consumer-grade NVIDIA Blackwell GPU and Apple Silicon (Metal) — no datacenter accelerator required.
A bounded field is not a universal upgrade. In some workloads it earns you nothing — and in those it costs you nothing either, so leaving it on does no harm. There is also one thing this run does not demonstrate. We'd rather you know the edges before a pilot than discover them after.
When the whole conversation fits comfortably inside the cache, there is nothing to bound — a standard cache carries it verbatim. SGF earns you nothing here, and costs you nothing: you can leave it on and lose nothing until the conversation outgrows the cache.
One question, one answer, no history to hold. The field is built for memory that accumulates; a prompt that never accumulates has nothing for it to preserve. No benefit, and no penalty.
The published run demonstrates topical coherence and continued attention to early-conversation content at depth. It does not demonstrate verbatim retrieval of a specific value from ~1.6 million tokens ago. If your workload turns on retrieving one exact string from deep history, that is not what this result shows. This run was not designed to test recall — exact retrieval is a separate capability that was not exercised here.
The published run uses a 1B model, chosen to stress the field rather than to flatter the output. Answer quality on that substrate is not the product and should not be read as a ceiling. It is a floor.
Everything a buyer needs to judge SGF is public: the footprint, the recall, the latency, and the conditions where it is the wrong tool. How centers are formed, scored, and merged is the invention, and it is held rather than published. The line is deliberate, and it does not move.
Footprint, recall, and latency, each with the model, hardware, context length, and lens configuration that produced it.
Where SGF underperforms: short contexts, single-turn prompts, and the edges this run does not demonstrate.
Parameter-level behavior, integration internals, and the pilot report measured on your own workload.
Center formation, merge scoring, and the pin policy, the mechanism behind every number.
A scoped four-week pilot: you bring one workload and a memory budget, we return a measured comparison against your current cache: footprint, recall, and the configuration that produced both.